What we’re doing

Getting what science knows about health to the people making real decisions, in a form they can actually use.

Discovery has no end and translation is the bottleneck. Why the loop between them is broken, where honest personal health information breaks down today, and what FRESH is building instead.

Information can be health. Most of what we know never arrives in a form anyone can act on, and that is a solvable problem.

Four people, four channels

A grandfather is closing in on eighty. He lives with his daughter and her family, three kids, three generations in one house, and whatever else the day has done to them, they eat dinner together. Around that table all four of them are being taught about health every day, and never by the same teacher.

The youngest is ten, and her teacher is the feed. It is gorgeous: rainbow cakes, foot-tall cookies, candy sculpted into art, because it is art, and none of it is food anyone needs. Between the cakes she is sold skincare in packaging she adores, full of ingredients she cannot pronounce and could not tell you the first thing about. Her brother is in his early teens, spent a couple of years heavier than he wanted to be, and loves video games more than almost anything. His teacher is nobody, because no channel anywhere is aimed at him, so whatever he comes to believe about food and his own body he will assemble alone out of whatever drifts past. Their mother learns from the wellness internet, which sells certainty by the serving: a banana a day for her period symptoms, this tea for that hormone, always one special food away from fixed.

The grandfather’s teacher, at least, went to school for this. He got a diagnosis and fifteen minutes: blood pressure, so less salt and less red meat. He grew up on Philly cheesesteaks and Sunday meatballs, with London broil when there was something worth celebrating, and none of the advice is wrong, exactly. It just arrives as a list of subtractions from the person he has been his whole life, with nothing in it about what a good week of eating could look like instead.

Four channels, then, shaped by different generations and different norms. Every one of them borrows its authority from the same place. The skincare ad gestures at studies, the banana post gestures at studies, and so does the doctor. The word they are all borrowing is science, which deserves some care, because science is less a source of truth than a way of earning answers slowly and in public. The best things anyone has ever known about health came out of that earning, which is exactly what makes the name worth borrowing for everything else.

Here is the strange part. Not one person at that table could tell you how the earning works. They trust science because school told them to, and the trust outlived the understanding.

That gap did not open by accident. The institutions that produce health knowledge answer to the next publication and the next grant, and nobody at this table appears anywhere in that accounting, so the trip from a finding to their kitchen is one that nobody is paid to make. The feed, meanwhile, answers to their attention every single day. One of those two systems is built to reach them and the other is not, and after a few generations of that, science has been designed out of daily life while the feed has been designed all the way in.

What a channel would have to get right

Suppose you wanted to build a channel that worked in their favour instead. Four things would have to be true at the same time, and each one fails in a way you can point at.

Accurate means the claim holds up against the evidence behind it, including the parts of that evidence that disagree, and that it is stated with the confidence the evidence actually supports rather than the confidence that makes it shareable.

Actionable means it names something a person could do this week, in their own kitchen, on their own budget. “Less salt” is a direction. “Here is what a good week of eating looks like for you, starting with the three things you already do well” is an action.

Accessible means it reaches people who do not already have a diagnosis, a referral, a degree, or money. Most of the good health information available today is gated behind at least one of those four.

Fun means somebody would choose it over whatever else is already open on their phone. Attention is the currency, and being worthy buys none of it.

The channel	Accurate	Actionable	Accessible	Fun
The ten-year-old’s feed	Rarely, and paid by the click	Instantly	Never off	It is the entire product
The mother’s wellness marketing	No, and accountable to no one	Very	Everywhere she looks	Yes
The grandfather’s fifteen minutes	Yes, for medicine	“Less salt” is not a plan	Only with a doctor and a diagnosis	No

The boy gets no row, because nothing is aimed at him at all, which is its own answer in all four columns. Nobody at that table gets all four at once, and the column where the good attempts usually die is the last one. We build on the assumption that people will tolerate boring if it is true, they decline, and we call that a motivation problem.

Access has one more gap in it, easy to miss: nobody in this family has ever sat down with a dietitian or a health coach, and it never occurred to any of them to try. Fifteen minutes has no room in it for a referral, physicians sit at the top of a ladder where everyone below them reads as less than expert, and if the one health professional in the room is not convinced that help with food and habit moves health, the referral will never occur to him either.

Now imagine the channels stay exactly where they are and only what flows through them changes. The ten-year-old follows creators her own age who are genuinely wise about health and every bit as fun as the cakes. Her brother finds that somebody finally built a channel that talks to him where he already spends his time. Their mother gets a way to read the quality of her family’s whole diet at once, so the weekly shop turns into the place she teaches her kids what she is learning. And the grandfather hears something nobody has thought to tell him: his fruit and vegetables are already enough, his blood pressure is now well controlled, and the best thing he can spend effort on now is moving more and building up his legs, because old age is a winter and he will want the strength to walk through it.

Same family, same table, same channels. Different information, every version of it adapted from the same source all four channels were already claiming. That is the whole project, and the rest of this page is about what it takes to build it honestly.

The family is not unusual

Nobody sits down to consult the informational environment. It arrives anyway, as the conversation at dinner, the answer you give a child who asks why one cereal is better than another, the headline you half-read on a phone, the thing a friend swears worked for them, the label you glance at with a cart in your hand. None of it announces itself as information, and by the time anyone makes a decision, belief has already taken shape in the background out of all of it. Nobody stands outside the information they grew up inside, experts included, which is why pointing at individual choices misreads the problem.

You can watch belief form ahead of the evidence in real time. Science reached its first international consensus on what gut health even means only recently, six domains, most of which the public has never heard of, and the public did not wait for it. In a survey published this month, 64 percent of Americans called gut health a high or extremely high priority, 54 percent already limit or avoid foods to protect it, and 84 percent rated their own gut health as good or better, going mostly on how they feel.¹ The same survey holds our family’s problem in a single pair of numbers: 74 percent name healthcare professionals as their most trusted source of information about gut health, and 12 percent have actually consulted one.

That gap would matter less if people were getting better at sorting claims on their own, and they are getting worse. In the 2023 international adult skills survey, 34 percent of US adults scored at or below the lowest measured level of numeracy against an OECD average of 25 percent, and the share at the bottom of the literacy scale went from 19 to 28 percent in six years.² American school results have been sliding since about 2013, with eighth-grade reading now at its lowest in the series and the largest share ever recorded below the basic level, while grades over the same period went up: average high school GPA rose from 3.22 to 3.39 in a study of 4.4 million students, and real transcripts show the same climb with no matching gain in twelfth-grade scores.³ The signal got louder as the thing it was measuring got quieter.

It is tempting to pin that decline on AI, but it began roughly a decade before ChatGPT existed. What the evidence does support is narrower and more useful. In the one randomized trial I know of, about a thousand high schoolers were split between a plain chatbot, a version built to give hints instead of answers, and no tool at all. The plain-chatbot group did 48 percent better at practice problems while they had it, then scored 17 percent below control on an exam once it was taken away. The hint-giving group was indistinguishable from control.⁴ What separated them was how the tool had been built.

That is this page’s argument aimed at a different target, because the same question can be asked of the system that produces health knowledge in the first place. How is that one built?

Discovery and delivery are one loop

I sat in an audience once while a room applauded an elder academic, and the applause was deserved. He had carried a compound the whole way, bench to pharmacy shelf, and he had the scars to show for it. Twenty years, start to finish. I clapped along with everyone else, and it took me a while to notice what we were celebrating: a twenty-year cycle, held up as the achievement of a career, with nobody in the room finding the number strange.

If anything the number is on the fast side. Clinical development alone, first dose in a human through to approval, runs a median of 8.3 years, and the longer arc from a promising laboratory finding to the clinical work that carries it has a median of 24.⁵ Most findings never make the trip at all: of 101 basic-science papers that promised a new treatment, five had reached licensed clinical use twenty-some years later, and one was in wide practice.⁶

The supply side has no such trouble. The National Library of Medicine took in more than 1.5 million new citations in a single year, carrying PubMed past 36 million,⁷ and the scientific literature as a whole doubles about every seventeen years.⁸ A great deal of that output competes with what already exists instead of adding to it. One registry lists 1,382 cardiovascular prediction models, 58 percent of which have never been tested on anyone outside the population that produced them.⁹ Systematic reviewers counted 78 nutrient profiling models carrying government backing through 2016, and 26 more by 2020, each of them deciding what counts as a healthy food.¹⁰ Some of the effort never reaches the science at all: at ten hours of formatting per paper and roughly five thousand papers a day, the field spends something like fifty thousand hours a day rewrapping its findings to suit one journal or the next.¹¹ Both of those inputs are estimates, though the Library’s own intake puts the daily count in that neighbourhood. All of it goes on the shape of the container, while the question of whether anything inside it reached a person goes unfunded and unasked.

Ask that question and the answer stays flat. Adults in the United States were getting about 55 percent of recommended care in 2003, and the 60-30-10 pattern (roughly 60 percent of care matching the evidence, 30 percent waste, 10 percent harm) has held for three decades.¹² Discovery compounds while delivery holds flat, because we keep opening new channels for finding things out and almost none for finding out whether the last thing worked.

Nobody chose that arrangement; it falls out of how the system is built. Map the organizations that sit between evidence and action and the stage almost nobody occupies is the last one, where you find out the thing is not working and stop. Stopping needs a rule set in advance, a rule needs to know whether the thing is working, and knowing that needs a signal of benefit. We built surveillance for harm and almost none for benefit, so nothing tells you it is working, so no stopping rule can be written, and nothing flows back into what gets studied next.

Without that return edge, the incentives that produced fifty competing dietary indices were never going to produce one useful answer, because nothing downstream was ever going to say which of the fifty was worth having. The argument over what counts as ultraprocessed is the current example. Enormous political and scientific capital goes into settling the definition, while the person holding a box in a grocery aisle gets no closer to knowing what to do, and no signal returns from that aisle to say so.

Nor is anybody positioned to fix that. Ask almost any researcher how it is going and you will hear some version of just surviving; the work gets produced and then goes nowhere in particular, piling up in the tens of thousands where no family will ever look. There has never been a real source of quality health information, or a system for delivering it, because there has never been an incentive to build one. Not in medicine, not in health care, not in the food industry, and not in prestigious academic institutions. That absence includes the classroom. There is no recognized field of study called health engineering, so we are not training anyone to conceive of this problem, let alone solve it. We train more tissue engineers, more physiologists, more statisticians, more ear, nose and throat doctors, more influencers, more politicians specializing in definitions of ultraprocessed food, all of them useful and some of them remarkable. Somewhere in the mix, somebody has to stop specializing and learn across the specialties, which was supposed to be the whole point of generalizable knowledge.

You can see how little the return trip is valued in the number the field uses to describe it. Everyone quotes seventeen years as the lag between evidence and practice. It comes from a chapter published in 2000 that assembled the figure from nine clinical areas, and one of its inputs was wrong by a factor of three, assuming that one in five people with diabetes got a foot exam when the real rate was closer to three in five. A review of 23 lag estimates found them ranging from under one year to more than 28, measuring different things at different points, and concluded that what we knew about lags was of limited use.¹³ The seventeen years gets repeated anyway, a quarter-century on. A field that has never gone back to revise its own measure of translation lag is telling you what it thinks translation is worth.

So the work is not to wait for certainty. Discovery never ends and should not slow down for anybody, and the honest position on almost any particular person is that most of what we would need to know about them is still unknown, and will be after another million papers. Anyone who tells you the science is settled at the level of your actual life is selling something. What has to change is the other half: translate what we have carefully, say how sure we are, and keep listening for what comes back. A study becomes a guideline, a guideline becomes something a person actually does, what happens next gets observed, and what comes back decides which question is worth asking next. Each half steers the other, and it is the second half, the carrying back, that has never been anybody’s job. That is the bottleneck, and it is a pity that delivery is still this bad.

How I got here

I was one of those tissue engineers. I spent years building multicellular experimental models of inflammation, regeneration, and disease, and somewhere in there I became convinced that the real problem was not our ability to generate health knowledge. It was our almost complete inability to steer that generation toward questions worth asking, and then to consolidate and deliver what comes back.

So I moved sideways, into nutritional epidemiology, and then into building. I have published in the field, worked with people who know far more than I do, and become steadily more humbled by how hard the cultural, data, and statistical problems actually are. FRESH is my attempt at building the missing half. It is not finished, I do not know that I am solving it correctly, and I would rather work on it in public than wait until I am sure.

What honest looks like in practice

Saying how sure you are is a design constraint, not a disclaimer bolted onto a finished product, and taking it seriously changes what can be built at all. Seven things follow from it, and together they are the part of this project I would defend hardest.

The doubt travels with the number. Measurement error is carried from the raw input through every step to the value on the screen, rather than dropped at the first convenient point and reintroduced as a footnote. A number without its uncertainty is a claim wearing the clothes of an answer, and it quietly takes the decision away from the person making it. If an estimate is shaky, you should be able to see that it is shaky and decide what to do about it. A wide interval is information in its own right, because it often says the honest move is to spend your effort somewhere the evidence is firmer.

Disagreement gets reported, not averaged. Quality never collapses to a single score here. When credible expert systems rate the same food differently, the spread is shown and the food is marked contested rather than quietly assigned a number. Averaging them produces a figure that no system actually endorses and hides the one fact the person most needs, which is that the experts have not settled it. Somebody confused about such a food is responding reasonably to a real disagreement, and they are owed the disagreement itself instead of a consensus invented for their comfort.

“Nobody knows” is a permitted answer. Sometimes a person brings a real problem and the honest response is that nobody knows yet. Not “here are five studies”, and not a confident guess dressed up in numbers. A system that can only speak when it has an answer goes quiet exactly when someone needs it most, and the silence gets filled by whoever is willing to sound certain. So the job is to say plainly that the question is unresolved, show what is known around the edges of it, say what evidence would settle it, and stay with the person while they decide anyway, because their decision does not wait for the literature. Most of the honest ground in health looks like that, and it is the part nearly every product skips.

The judgment inside the model is shown. Science is not unbiased by nature, whatever the comfortable version of it says. Every risk score, index, and algorithm has human choices inside it: which population it was fitted on, what counted as an outcome, what was treated as noise, where a threshold was drawn. Those choices move the answer, sometimes more than the data does. Bayesian methods are the right home for this work because they force the assumptions into the open where somebody can argue with them. Hiding the judgment is how trust and understanding came apart in the first place, and no version of rebuilding that trust begins by hiding it again.

Every claim keeps a trail back to its source. The footnotes on this page are the standard, and they include the parts that cut against the argument, like the caveats about how the adult skills survey changed between cycles, or the school results that went up rather than down. Reporting only the supporting half is the most common dishonesty in this field, and it stays invisible unless somebody shows their work. If a claim here cannot be walked back to something you can read yourself, it does not belong here.

Nothing counts as working until it reproduces a published benchmark. Standard error included, not just the point estimate. A method that hits a published mean while missing its uncertainty has agreed by accident, and it will fail the first time it meets data the original authors never saw. Holding to this is slow and it kills features. It is also the only claim of correctness worth anything.

What people believe gets measured as seriously as what food contains. Beliefs are treated as data here, measured and then compared against the evidence, which is the signal the loop has been missing. The comparison comes back specific: foods the public judges more harshly than the evidence supports, foods it lets off too easily, and foods the experts cannot agree on either. Naming which is which, food by food, is an education agenda rather than a scolding, and it decides what is worth saying next by finding out what people actually think instead of assuming.

None of that is what the platform is for. It is what makes the platform worth trusting, and it is a better position to argue from than false confidence. A person deciding what to cook tonight can work with “we are fairly sure about this, much less sure about that, and here is what would change our minds”. Nobody can work for long with confidence that turns out to have been borrowed, which is the thing already failing them at the dinner table.

Food first, then the rest of a life

Food comes first because it is the hardest case and the most daily one: expert systems clash most openly there, measurement is noisiest, and public belief has drifted furthest from the evidence. The methods themselves are not about food. The same machinery reaches sleep, movement, mind and mood, social connection and loneliness, self-care, substance use, biological age, health literacy, and the place you live, all read from what you report, what you wear, and what your blood says.

What it is made of

Five engines sit behind the tools and essays here.

Engine	What it does
Evidence	Reads the published literature end to end: finds the studies, checks trial quality and trustworthiness, re-pools the effects, and keeps a trail from every claim back to its source.
Consensus	Reconciles competing expert rating systems into one answer with the disagreement measured and reported, never hidden inside an average.
Risk	Combines risk models with trial-derived effects to estimate what a change is actually worth for particular people and places, with uncertainty propagated rather than dropped.
Decision	Turns contradictory signals into one personal answer that shows its own doubt.
Well-being	Scores well-being across an open set of life domains from self-report, wearable, and biomarker inputs, per domain and combined.

Together they are meant to do what no single adviser can, which is the framing set out on the front page.

Who this is for

Anyone who wants to see the evidence for themselves: people making decisions about their own health, the people who care for them, and the people writing the labels, the guidance, and the policy. The essays are the argument in public. The tools let you check it.

If you are thinking about health measurement, clinical risk, nutrition science, science reform, or what honest evidence synthesis even looks like, I want to hear from you. So let’s start talking better, asking better questions, and building tools that turn scientific complexity into better health decisions.

Footnotes

1. International Food Information Council, Spotlight Survey: Americans’ Perceptions of Gut Health, published 4 August 2026; fielded online 20–25 May 2026 among 1,004 US adults aged 18 and over, weighted for proportionality. Figures quoted here are from IFIC’s accompanying release: gut health a high or extremely high priority, 64 percent; limit or avoid foods or beverages to support it, 54 percent; own gut health rated good, very good or excellent over the past 30 days, 84 percent; healthcare professionals named a trusted source, 74 percent; healthcare professional consultation among actions taken in the past year to support gut health, 12 percent. Two things a reader should weigh: IFIC is funded by the food and beverage industry, and the 74 percent and 12 percent measure different things, trust in a source against an action taken in one year, so the pair is a fair illustration rather than a strict contradiction. The six domains of gut health come from the first international consensus statement on the subject, published by the International Scientific Association for Probiotics and Prebiotics.

2. OECD, Survey of Adult Skills 2023, United States country note, and NCES, Highlights of the 2023 U.S. PIAAC Results, NCES 2024-202. US adults averaged 249 in numeracy, below the OECD average; 34 percent scored at or below Level 1 against an OECD average of 25 percent; the at-or-below- Level-1 share rose from 29 to 34 percent in numeracy and 19 to 28 percent in literacy between 2017 and 2023. The overall response rate was 28 percent, and NCES warns that cross-cycle comparisons need care because the framework was revised, administration moved to tablet-only, and basic-skills items counted toward overall scores only in 2023, which mechanically moves some adults down a level. Both caveats cut against the trend I am describing, so treat the direction as better supported than the exact size.

3. Eighth-grade reading peaked around 2013 and has fallen every cycle since, reaching 258 in 2024, five points below 2019, with about a third of eighth graders below the NAEP Basic level, the largest share recorded: NAEP reading 2024. Twelfth-grade reading in 2024 was the lowest in the 32-year trend. Not everything is falling: fourth-grade math rose 2 points in 2024, and nine-year-olds gained 4 points in both subjects on the 2025 long-term trend, which is worth saying because leaving it out would flatter the argument. On grades, Sanchez & Moore, Grade Inflation Continues to Grow in the Past Decade, ACT Research Report R2134, 2022: 4,393,119 students, average high school GPA 3.22 in 2010 to 3.39 in 2021, and still 3.17 to 3.36 after adjusting for student and school characteristics. GPA there is self-reported and the ACT-tested population changed composition over the period, so the transcript-based federal figure is the firmer one: the NAEP High School Transcript Study recorded graduate GPA rising from 3.0 in 2009 to 3.11 in 2019 with no matching gain in twelfth-grade scores. NAEP is descriptive and identifies no causes.

4. Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: evidence from high school mathematics,” PNAS 2025;122(26):e2422633122, PMC12232635. Roughly 1,000 students across about 50 classrooms in one Turkish high school, randomized to control, a ChatGPT-like assistant, or a tutor version giving hints rather than answers. With access: 48 percent and 127 percent improvements. On a later exam without access: the unguarded group scored 17 percent below control, the tutor group was indistinguishable from it. One school, one subject, short-term outcomes only, by the authors’ own statement. I am citing this trial rather than the widely shared MIT EEG essay-writing study, which is an unreviewed preprint of 54 people (18 in the session carrying its main claim) whose authors have publicly asked that it not be described as showing harm or damage, or the Microsoft and Carnegie Mellon survey of 319 workers, which measures what people believe about their own effort rather than what they can do.

5. Two different clocks, and it matters which one you mean. The 8.3 years is a median clinical development time, first-in-human dosing to regulatory authorization, for innovative drugs approved by the FDA over the preceding decade: Brown, Wobst, Kapoor, Kenna & Southall, “Clinical development times for innovative drugs,” Nature Reviews Drug Discovery 2022;21(11):793–794, PMC9869766. The 24 years is a median (IQR 14–44) from the first description of a promising basic-science finding to the earliest highly cited clinical article built on it: Contopoulos-Ioannidis, Alexiou, Gouvias & Ioannidis, “Life cycle of translational research for medical interventions,” Science 2008;321(5894):1298–1299, PMID 18772421. Neither figure is “twenty years from bench to shelf” as a career is usually described, and I have not found a clean published measure of that. The applause in the story was for one person’s account of their own working life, not for a statistic.

6. Contopoulos-Ioannidis, Ntzani & Ioannidis, “Translation of highly promising basic science research into clinical applications,” American Journal of Medicine 2003;114(6):477–484, PMID 12731504. Of 101 articles published 1979–1983 making an explicit novel therapeutic or preventive promise, 27 led to a published randomized trial, 19 produced a favourable trial, five reached licensed clinical use, and one was in extensive clinical use for the licensed indication at the time of follow-up. A single cohort from a single window, and the denominator is promises rather than papers.

7. National Library of Medicine, MEDLINE/PubMed production statistics. For fiscal 2023, the most recent year the page reports: 1,567,478 citations added to PubMed, cumulative total 36,555,430. PubMed’s own front page now describes the catalogue as more than 40 million, which is the later and less precisely dated figure, so the sourced number is used here. Annual additions include records entered retrospectively, so they are not a count of articles published that year.

8. Bornmann, Haunschild & Mutz, “Growth rates of modern science,” Humanities and Social Sciences Communications 2021;8:224, doi:10.1057/s41599-021-00903-w. Overall growth of 4.10 percent a year, doubling time 17.3 years, across Dimensions, Microsoft Academic, Web of Science and Scopus, 1900–2018. All of science, not biomedicine alone.

9. Wessler, Nelson, Park, et al., “External Validations of Cardiovascular Clinical Prediction Models,” Circulation: Cardiovascular Quality and Outcomes 2021;14(8):e007858, PMC8366535. The Tufts PACE registry held 1,382 cardiovascular models published 1990 through March 2015; 807 of them, 58 percent, had never been externally validated. Among those that were, median discrimination fell 11.1 percent from the derivation estimate. The cut-off is 2015, so the count today is higher, not lower.

10. Labonté, Poon, Gladanac, et al., “Nutrient Profile Models with Applications in Government-Led Nutrition Policies,” Advances in Nutrition 2018;9(6):741–788, PMC6247226, which included 78 models after screening 387 candidates; and Martin, Turcotte, Cauchon, et al., an update covering May 2016 to September 2020, Advances in Nutrition 2023;14(6):1499–1522, PMC10721541, which added 26. Both count only models developed or endorsed by governmental and intergovernmental bodies, so roughly a hundred is a floor and not a census. The commercial and proprietary ecosystem is not systematically catalogued by anyone, and counting it is what takes the total past two hundred, the figure used elsewhere on this site. This page uses the government-backed count because it is the part with a citable denominator, which means it understates the problem rather than overstating it.

11. My own arithmetic, first posted on LinkedIn: ten hours of formatting per paper times five thousand papers a day. Both inputs are estimates rather than measurements. The paper count is the defensible half, since NLM’s fiscal 2023 intake of 1,567,478 citations works out to roughly 4,300 a day; the ten hours is a plausible figure I have not seen measured, so treat the product as an order of magnitude and not a finding.

12. McGlynn, Asch, Adams, et al., “The quality of health care delivered to adults in the United States,” New England Journal of Medicine 2003;348(26):2635–2645, PMID 12826639: participants received 54.9 percent (95% CI 54.3–55.5) of recommended care across 439 quality indicators in 12 metropolitan areas. The 60-30-10 framing is Braithwaite, Glasziou & Westbrook, “The three numbers you need to know about healthcare,” BMC Medicine 2020;18:102, PMC7197142, who describe the pattern as persisting for three decades. The 2003 measurement is old, which is part of the point, and also a limitation: nobody has repeated it at that scale.

13. The primary source is Balas & Boren, “Managing clinical knowledge for health care improvement,” Yearbook of Medical Informatics 2000:65–70, doi:10.1055/s-0038-1637943, which measured time to 50 percent clinical use across nine clinical areas and added a publication lag to reach seventeen years. The factor-of-three error is documented in Howard, “Evidenced-Based Claims About Evidence,” MDM Policy & Practice 2017;2(2), PMC6125044: the analysis assumed 20 percent of patients with diabetes received foot exams in 1998 when the CDC rate was closer to 60 percent. The review of 23 estimates is Morris, Wooding & Grant, “The answer is 17 years, what is the question,” Journal of the Royal Society of Medicine 2011;104(12):510–520, PMC3241518, which found lags from under 1 year to over 28 and concluded that existing knowledge of them is “of limited use.” I am using the number as an exhibit rather than as evidence, which is the only honest way to use it.